564949 - Emit new opcodes that predict free variables as being on the global object

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Anderson [:dvander] - inactive, e-mail if emergency]

Free variable references that are not polluted by eval/with, and cannot otherwise be shadowed by anything on the scope chain, will always be a ReferenceError or resolved on the global object.

We would like to create new opcodes, or annotations for the existing *GLOBAL* opcodes, so the method JIT can create monomorphic inline caches for such free references.

Comment 1

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: wip

removes gvar ops. adds new GNAME ops which act like NAME ops, except the object is predicted to be the global object. this is forced both in the interpreter and tracer for good measure, though is technically not needed.

a testament to our def-use chains, all this patch really does is rewrite NAME ops in BindNameToSlot, and passes reftests and trace-tests. this might not be the right approach, but it's enough to start adding monomorphic PICs to JM.

Comment 2

[David Anderson [:dvander] - inactive, e-mail if emergency]

Comment on attachment 445280 [details] [diff] [review]
wip

About to try monomorphic ICs on JM. Brendan, any thoughts on whether these parser changes are on the right track?

Comment 3

[Brendan Eich [:brendan]]

Is this bug still valid?

/be

Comment 4

[David Anderson [:dvander] - inactive, e-mail if emergency]

(In reply to comment #3)

It landed on JM a while ago as: http://hg.mozilla.org/users/danderson_mozilla.com/moo/rev/986b89529822

I'd still appreciate your feedback but I think it can be marked as FIXED, no big motivation to pre-land on TM with JM landing in sight.